◆Healthcare
Retrieval-augmented generation (RAG) is how you get Claude to answer from your reality — your policies, guidelines, and clinical content — instead of its general training. In healthcare it's the difference between a plausible answer and a defensible one. The complication is PHI, and it changes how the system must be built.
Instead of relying on what a model memorized, RAG retrieves the relevant passages from your own documents at question time and asks Claude to answer using only those passages, with citations. The result is grounded, current, and traceable: a prior-authorization rationale that points to the exact policy clause, or a care-operations answer that links to the source SOP. That traceability is exactly what reviewers and auditors need, and it sharply reduces the unsupported "confident guess" that makes generic AI unusable in clinical settings.
The moment protected health information is involved, retrieval quality is necessary but not sufficient. The architecture has to satisfy privacy and security obligations from the first line of code:
Your security officer will evaluate the system against the same Security Rule safeguards they apply to any system touching ePHI, so it pays to speak that language from the first architecture review. Access control (45 CFR 164.312(a)) maps to unique identities for every user and service, role-scoped retrieval, and automatic session controls. Audit controls (164.312(b)) map to the immutable log of who asked what, which chunks were retrieved, and what was returned. Transmission security (164.312(e)) maps to encryption in transit across every hop — including the hop to the model API. None of these are exotic; what's new is making sure the retrieval layer inherits them rather than quietly bypassing them.
Two structural points deserve early attention. The minimum-necessary standard (164.502(b)) argues for passage-level retrieval rather than whole-chart dumps — a design choice that also happens to improve answer quality. And the business-associate chain must be complete: your organization, the implementation partner, the model provider, and any vector-database or hosting vendor in between. A RAG system is only as compliant as the least-papered vendor in its data path.
The fastest deployments often start where PHI isn't needed. A policy-and-guidelines assistant — formularies, medical policies, SOPs, payer manuals — delivers daily value to care-operations and support teams with no patient data in the corpus at all. When patient context is genuinely required, consider whether de-identified data is enough for the use case: HIPAA recognizes Safe Harbor (remove the eighteen identifier categories, 164.514(b)(2)) and Expert Determination (164.514(b)(1)) as paths to data that is no longer PHI. Each PHI-free workflow you ship builds the operational evidence that makes the PHI-bearing ones easier to approve.
A handful of engineering decisions do most of the compliance work. The index must be permission-aware: retrieval respects the same access controls as the source systems, so a user can never surface a document through the assistant that they couldn't open directly. Every chunk carries provenance — source document, section, version, effective date — so citations point at something real and answers built on superseded policy are detectable. Application logs and telemetry are scrubbed of PHI by design, because debug logs are where private data most often leaks. And the index itself is governed: when a source document is corrected or deleted, the change propagates on a defined schedule, and retention follows your records policy rather than defaulting to forever.
RAG should inform a decision, not make it. For prior authorization, claims, or any clinical determination, the system drafts a recommendation with its supporting evidence and a qualified human approves it. That keeps accountability where it belongs and keeps the workflow inside the lines regulators and clinical leaders expect.
Before launch, build a golden set of real questions with clinician-agreed correct answers, and measure three things: faithfulness (does the answer say only what the cited passages support?), citation coverage (does every substantive claim carry one?), and abstention (when the corpus doesn't contain the answer, does the system say so instead of improvising?). The right behavior for an unanswerable question is "I can't find this in policy" plus a pointer to a human — and that behavior should be tested as deliberately as the happy path. After launch, a scheduled sample of live answers goes to clinical review, so quality drift is caught by process rather than by incident.
Every answer should link back to the specific passage it came from. Citations turn the system from a black box into something a reviewer can verify in seconds, an auditor can trace, and a clinician can trust. If an answer can't cite its source, it shouldn't be shown.
Done this way, RAG gives healthcare teams speed without giving up control — grounded, cited, minimized, permission-aware, and overseen. That's the difference between a chatbot near PHI and clinical-grade infrastructure.